through Symbolic Regression*

Vanadium dioxide (VO₂), a correlated semiconductor, can exhibit a metal-insulator transition (MIT). With suitable choices of dopants on either the cation or anion site, it is possible to tune the IMT temperature by several tens of degrees and control the hysteresis. A challenge in modelling the effects of control parameters, such as doping concentration, type of dopants, etc, on the IMT is the complexity associated with experimental procedures. The scarcity of experimental data hinders the development of modern Artificial Intelligence (AI)/Machine Learning (ML) models, with only a few empirical linear models available. A promising approach for bridging the gap between physical reasoning and data-driven methodologies is the Symbolic Regression (SR) approach. SR can identify nonlinear analytical expressions connecting target properties to the key input parameters, even with relatively small datasets. In this work, we develop SR models to capture the trends in the IMT in VO₂ affected by different dopant control parameters. We use a comprehensive set of experimental data from a wide range of literature sources reported over the past two decades. Our study reveals a dual nature of the IMT transition for different Tungsten (W) doping concentrations likely due to electronic (low doping) versus structural interactions (dominant at high doping densities). Our models can further uncover mechanistic insights on controlling phase transitions useful for realizing low-power Mott electronic devices.